RE: Non-packed structures in IP headers

[David Laight <David.Laight@ACULAB.COM>]

From: Florian Weimer
> Sent: 01 October 2021 21:10
> 
> * Carles Cufi:
> 
> > I was looking through the structures for IPv{4,6} packet headers and
> > noticed that several of those that seem to be used to parse a packet
> > directly from the wire are not declared as packed. This surprised me
> > because, although I did find that provisions are made so that the
> > alignment of the structure, it is still technically possible for the
> > compiler to inject padding bytes inside those structures, since AFAIK
> > the C standard makes no guarantees about padding unless it's
> > instructed to pack the structure.
> 
> The C standards do not make such guarantees, but the platform ABI
> standards describe struct layout and ensure that there is no padding.
> Linux relies on that not just for networking, but also for the userspace
> ABI, support for separately compiled kernel modules, and in other
> places.

In particular structures are used to map hardware register blocks.

> Sometimes there are alignment concerns in the way these structs are
> used, but I believe the kernel generally controls placement of the data
> that is being worked on, so that does not matter, either.
> 
> Therefore, I do not believe this is an actual problem.

And adding __packed forces the compiler to do byte accesses
(with shifts) on cpu that don't support misaligned memory accesses.

So it really is wrong to specify __packed unless the structure
can be unaligned in memory, or has a 'broken' definition
that has fields that aren't 'naturally aligned'.
In the latter case it is enough to mark the field that requires
the padding before it removed as (IIRC) __aligned(1).
The compiler will then remove the padding but still assume the
field is partially aligned - so my do two 32bit access instead
of 8 8bit ones).

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)